BUG=http://b/270223278
Add convolution ops into the requantizer tool. Since conv ops are quantized per-channel, per-channel requantization on biases are also introduced in this PR.
BUG= http://b/268089648
Fix issues for the requantizer when the model has constant tensors. Some ops (like ADD) may operate on constant tensors, which require them to be requantized to int16 as well.
Also added additional ops for layer norm
BUG=http://b/266447344
A flatbuffer manipulation tool to requantize a int8 activation, int8 weight (8x8) PTQ model to int16 activation, int 8 weight (16x8) PTQ model. This tool targets for LSTM models only since the 16x8 LSTM is not supported in the standard TFLite conversion process.
